accelerate
Avoid duplicating memory for tied weights in `dispatch_model`, and in forward with offloading
#2330

Merged

Commits

wip

invalid-email-address committed 2 years ago
fix

invalid-email-address committed 2 years ago
add test

invalid-email-address committed 2 years ago
cleanup

invalid-email-address committed 2 years ago
style

invalid-email-address committed 2 years ago
style & tests pass

invalid-email-address committed 2 years ago
fix offload, submodules

invalid-email-address committed 2 years ago
cleanup

invalid-email-address committed 2 years ago
Update tests/test_big_modeling.py

fxmarty committed 2 years ago
Update tests/test_big_modeling.py

fxmarty committed 2 years ago
disk offloading do not reload tied parameters in memory

invalid-email-address committed 2 years ago
remove outdated comment

invalid-email-address committed 2 years ago